Machine/Deep Learning Based Glacier Surface Segmentation of ISRO-ASAR Datasets

With increasing digital image data obtained from aircraft and satellites, remote sensing image analysis is becoming crucial in cryospheric studies. A new dataset was acquired by ISRO-ASAR radar over the western United States in December 2019. Synthetic Aperture Radar (SAR) data, in general, is underutilized but highly useful. With the NISAR mission launching in 2023 and hosting sensors that are similar to the ISRO-ASAR test instrument, it is crucial to explore the test data. We have utilized multi-frequency L-band and S-band ISRO-ASAR covering the Chugach mountains and southeast Fairbanks in Alaska. We are employing state-of-the-art machine learning and deep learning algorithms, targeting object detection and segmentation, such as Feature Pyramid Networks (FPN), SegNet, and DetNet, to show the potential ability of the datasets for glacier studies, including glacier surface segmentation. The datasets were pre-processed and augmented before being fed into the networks to prevent overfitting and provide better model generalization. Our preliminary results show the unique capabilities of L- and S-band data in glacier surface segmentation, especially when utilizing cross-polarized returns. Further experiments are being conducted to learn the performances and unique benefits of L-band and S-band and their corresponding polarization for machine/deep learning applications.